RTX 4090 (AD102): dmaAllocMapping_GM107 VA space exhaustion → RC watchdog GPU lock after 7 days uptime (nvidia-open 595.58.03)

RTX 4090 (AD102): dmaAllocMapping_GM107 VA space exhaustion → RC watchdog GPU lock after 7 days uptime (nvidia-open 595.58.03)

System Information

Bug Description

After approximately 7 days of uptime under normal desktop use (Wayland compositor, browser, terminals), the nvidia-open driver began failing DMA VA space allocations, eventually leading to a complete GPU lockup that required a hard reboot to recover.

Timeline

1. May 3 12:38 — System boot. nvidia-open 595.58.03 loads normally.
2. May 10 13:42 — First burst of 19× dmaAllocMapping_GM107: can't alloc VA space for mapping errors in rapid succession.
3. May 10 14:44–14:50 — Three more bursts of DMA allocation failures (22, 15, 17, 19 occurrences).
4. May 10 17:24 — Hyprland crashes (coredump in CCompositor::initServer). RC watchdog fires:

   NVRM: krcWatchdog_IMPL: RC watchdog: GPU is probably locked!  Notify Timeout Seconds: 7
   NVRM: nvAssertFailedNoLog: Assertion failed: GPPut < WATCHDOG_GPFIFO_ENTRIES @ kernel_rc_watchdog.c:1549
   

5. May 10 17:24–17:35 — SDDM attempts to restart Hyprland 3 times; each attempt results in the same crash at CCompositor::initServer. Display server fails to start on vt 2.
6. May 10 17:36 — Manual attempts to reload the nvidia modules via modprobe -r nvidia_drm nvidia_modeset nvidia_uvm nvidia fail — GPU is hung.
7. May 10 17:36–17:38 — RC watchdog continues firing every ~8 seconds until system reboot. The GPU never recovers.
8. May 10 17:38 — System rebooted. GPU functions normally on the new boot, no errors present.

Logs

DMA allocation failures (first burst):

May 10 13:42:15 kernel: NVRM: dmaAllocMapping_GM107: can't alloc VA space for mapping.
[repeated 19 times]

RC watchdog (repeating every ~8s from 17:36 until reboot):

May 10 17:36:04 kernel: NVRM: krcWatchdog_IMPL: RC watchdog: GPU is probably locked!  Notify Timeout Seconds: 7
May 10 17:36:05 kernel: NVRM: nvAssertFailedNoLog: Assertion failed: GPPut < WATCHDOG_GPFIFO_ENTRIES @ kernel_rc_watchdog.c:1549

Hyprland coredump (all 3 restart attempts identical):

Process (Hyprland) of user 1000 dumped core.
#10 /usr/bin/Hyprland + 0x15d0af
#11 CCompositor::initServer (/usr/bin/Hyprland + 0x2ac4c0)
#12 main (/usr/bin/Hyprland + 0x206c53)

PCI BAR layout:

BAR0: 0xfb000000–0xfbffffff  (16 MB, 32-bit non-prefetchable)
BAR1: 0xd0000000–0xdfffffff  (256 MB, 64-bit prefetchable)
BAR3: 0xe0000000–0xe1ffffff  (32 MB, 64-bit prefetchable)

Notes

• No Xid errors were logged.
• The system has 24 GB VRAM. Workload at the time of failure was a normal desktop session — not a heavy VRAM consumer.
• The proprietary driver was not tested.
• SSH remained functional after display loss, suggesting the host CPU/kernel was not frozen — only the GPU was locked.
• A separate set of AMD MCE hardware errors (bogus, all-0xffffffffffffffff MC31_STATUS values from an unrelated kernel regression) were present throughout the boot but are not believed to be related to the GPU failure.

Related Issues

• #1078 — Same driver (595.58.03), same dmaAllocMapping_GM107 errors on RTX 4060 Ti
• #1132 — Same symptom chain (dmaAllocMapping → krcWatchdog GPU lock) on RTX 5070 / 595.71.05
• #1134 — BAR mapping VA space exhaustion on RTX 3090 / 595.71.05
• #971 — GPU hang after extended uptime (~66 days) on B200 / 570.133.20

This appears to be part of a regression in the 580–595 open driver series affecting DMA VA space management across multiple GPU generations.

Start by reviewing the DMA allocation failures and RC watchdog logs, then read kernel_rc_watchdog.c around line 1549 and compare related issues #1078, #1132, and #1134. Done would require identifying the DMA VA exhaustion cause and validating a driver-level fix against the long-uptime failure sequence.